Using Optical Flow Fields for Polyp Detection in
Virtual Colonoscopy
Burak Acar1, Sandy Napel1, David Paik1, Burak Go¨ktu¨rk2, Carlo Tomasi3,
and Christopher F. Beaulieu1
1 Stanford University, Department of Radiology
LUCAS MRS Center, 3D Lab., Stanford, CA 94305-5488, USA
bacar, snapel@stanford.edu
paik@smi.stanford.edu, cfb@s-word.stanford.edu
www.3dradiology.org
2 Stanford University, Electrical Engineering Department
Stanford, CA 94305, USA
3 Stanford University, Computer Science Department
Stanford, CA 94305, USA
Abstract. Since the introduction of Computed Tomographic Colonog-
raphy (CTC), research has mainly focused on visualization and naviga-
tion techniques. Recently, efforts have shifted towards computer aided
detection (CAD) of polyps. We propose a new approach to CAD in CT
images that attempts to model the way a radiologist recognizes a polyp
using optical flow fields (OFF). Features extracted from OFFs are used
by a linear classifier for polyp detection. An initial validation of our tech-
nique resulted in an average of 75% specificity at 100% sensitivity in a
10-fold cross validation study on a set of 220 polyp-like structures, 20 of
which were true polyps.
1 Introduction
Computed Tomographic Colonography (CTC) was suggested in the early 1980’s
and realized in the 1990’s as a minimally invasive method that would make mass
screening of colorectal cancer feasible [1,2,3,4]. Since then several studies were
conducted to assess the performance of CTC [5,6,7,8,9,10]. In almost all studies,
CTC was based on the examination of CT images by an expert radiologist, using
either the 2D images, 3D virtual colonoscopic views or both. Thus, until recently
most efforts were directed towards developing better visualization and navigation
techniques [4,11,12,13,14,15,16,17,18,19]. Lately, more effort has been put into
computer aided detection (CAD), whose ultimate goal is to identify polyps in a
3D CT data efficiently, and with high sensitivity and specificity. Mir et al. reviews
a set of techniques that can be used for shape description in CT images [20].
We are aware of two studies on CAD in CTC: Summers et al. reported a shape
based polyp detector and concluded that CAD is feasible in CTC [21] and Paik
et al. proposed a Hough Transform-based polyp detector [22,23]. However, these
CADs suffer from low specificity, which would require radiologists to examine a
large number of CAD hits to rule out false detections.
W. Niessen and M. Viergever (Eds.): MICCAI 2001, LNCS 2208, pp. 637–644, 2001.
c
© Springer-Verlag Berlin Heidelberg 2001

638 B. Acar et al.
The goal of our research is to develop a highly sensitive and specific CAD
for CTC. We attempted to model the way a radiologist recognizes a polyp by
focusing on the changes in consecutive images as one views sequential cross-
sections of the volumetric source CT data (3D CT data). We used Optical Flow
Fields (OFF) to represent these changes, and a linear classifier acting on these
features to recognize true polyps.
2 Method
2.1 Data Acquisition
Patients were imaged in the supine position after colon cleansing and air-insuffla-
tion of the colon on a GE HiSpeed Advantage single detector or GE LightSpeed
multi-row detector scanner (GE Medical Systems, Milwaukee, WI). Typical ac-
quisition parameters were 3 mm collimation, pitch 1.5-2.0, 1.5 mm reconstruc-
tion interval, 120 KVp, 200 MAs for the single detector scanner and 2.5 mm
collimation, pitch 3.0, 1.0-1.5 mm reconstruction interval, 120 KVp, 56 MAs
for the multi-row detector scanner. Data were stored as 2 byte unsigned in-
tegers. The size of the 3D data matrix was 512 × 512 × N , N is the num-
ber of axial slices. N is typically around 350. The average voxel spacing was
0.74mm× 0.74mm× 1.31mm.
2.2 Initial Detection
The data was preprocessed by a Hough Transform-based polyp detector (HTD)
[22,23]. HTD basically computes the normals to an isointensity surface and
searches for voxels at which a large number of normals intersect. Defining the
number of normals that intersect in a given voxel as the HT score of that voxel, a
threshold is applied for detection. Typically the voxels at the centers of spherical
structures have high scores. The detected voxels mark the locations of polyps
with high sensitivity but low specificity. We extracted a subvolume consisting
of 21 × 21 × 21 voxels centered on each detection with a score above a fixed
threshold.
2.3 Optical Flow Field Computation
The aim of the OFF computation is to characterize the change in the location
of the edges (tissue/air boundary) in the image plane while scrolling back and
forth along the third dimension. This third dimension can be thought as the
time axis [24]. If we name the image plane as the xy-plane then the basic optical
flow equation is [25]:
∇I.v + ∂I
∂t
= 0 (1)
where v(x, y) is the OFF and I(x, y) is the image, i.e. the intensity function.
Equation 1 allows only the computation of v
⊥
, the component along the local